View
1
Download
0
Category
Preview:
Citation preview
Roads, Trade and Urbanization in the Tropics: Theory
and Evidence from the Brazilian Amazon ∗
Heitor S. Pellegrina
Brown University
PRELIMINARY VERSION
August 5, 2014
Abstract
The Brazilian Amazon has undergone a great transformation of its economy since theend of the 1990s, when the price of its export-oriented commodities increased in morethan 200% in a few years. However, counties in the region experienced the effects of thissoaring price differently according to their access to a federal road system constructedin the early 1970s. I take advantage of this institutional setting, and a rich datasetwith annual information on agriculture production and labor market employment, tostudy how trade affects urbanization in developing countries. I find that countieswith access to roads expanded more their export of capital intensive commodities andtheir employment in urban activities after the price shock. Furthermore, structuralestimates of a model indicate that an urbanization process would not have occurred inthe absence of a capital intensive activity. This article provides evidence that the lackof good trade opportunities in terms of price and products may undermine the benefitsof transportation projects in developing countries.
Keywords: Urbanization, Transportation, Trade, Commodities
JEL Classification: F16, O13, O18, O54, N56, Q11, R23
∗Email: heitor pellegrina@brown.edu. I am grateful to my advisers Andrew Foster, Jonathan Eaton,Sriniketh Nagavarapu, and Nathaniel Baum-Snow. I am also thankful to Vernon Henderson, Marcos Rangel,Lucas Scottini, Alex Eble, Ken Chay, David Weil, Leah VanWey, Alexei Abrahams, and participants atpresentations held at the Econ department at Brown, the PSTC, the University of Sao Paulo, the WatsonInstitute for International Studies, and the Readings and Research group on Trade. I am specially thankfulto Peter Richards for many helpful comments and for sharing his vast knowledge of the Amazon. Data forcommodity prices in Mato Grosso was provided by IMEA. This research was supported by the NationalScience Foundation’s IGERT and by the Institute for the Study of Environment and Society.
1
1 Introduction
Urbanization is one of the most established symptoms of development (Kuznets, 1973; Her-
rendorf et al., 2013). As countries experience economic growth, there is a reallocation of
labor from rural activities to urban ones. Theoretically, most of the studies explain the
process of urbanization using closed economy models, where a combination of technological
improvements in agriculture and a inelastic demand for food decreases the need for labor
in the rural sector. Empirically, this literature provides stylized facts at the country level:
either comparing the level of urbanization of different countries at a given period, or the
same country over time. This body of research, however, still provides very limited microe-
conomic evidence of the theoretical mechanisms that explain when and why urbanization
occurs 1, and, also, how international trade may affect this process (Foster and Rosenzweig,
2007; Gollin et al., 2013; Henderson et al., 2013).
In this article, I work on these two limitations of the current literature by investigating
the microeconomic evidence on the expansion of trade and urban activities in the Brazilian
Amazon. This region has undergone an intense expansion of its trade activities since the
end of the 1990s, when the price of commodities in terms of the local currency increased
more than 200% due to variations in the Brazilian exchange rate and the global increase in
commodity prices. However, municıpios (counties) experienced the effects of these macro
economic changes differently according to their access to federal roads. The federal highway
system in the Amazon was constructed in the early 1970s, and ended up providing the main
arterial routes for exporting goods when the commodity prices soared in the end of the 1990s.
I explore this institutional setting to study how access to transportation leads to differences
in the expansion of trade and urban activities when macro-economic conditions for export
improve.
To guide the empirical work, I build a small open economy model with a set of counties
with heterogeneous distance to federal roads. The model leads to an expansion of urban
activities through the consumption of non-tradeable goods, which are assumed to be pro-
duced by the urban centers in each county. Counties that engage in trade activities increase
their income, which translates into a higher demand for both tradeable and non-tradeable
goods (a Balassa-Samuelson effect). Because non-tradeable goods can only be provided by
local labor, the increase in demand pushes wages up in the county, which brings migrants
from other regions of Brazil to work in the non-tradeable (urban) sector. This in-migration
1See Fajgelbaum and Redding (2014); Storeygard (2013); Jedwab (2013) for recent articles exploring themicroeconomic evidence on urbanization.
2
then reduces wages in the county, leading the local economy to an equilibrium. Based on the
characteristics of the Amazon economy, I include the presence of a specific activity with large
fixed costs for machine acquisition. Farmers with better access to federal roads face higher
output prices because they pay lower transportation costs, and, as a consequence, they are
more likely to have profits that are large enough to compensate the fixed costs of venturing
in the production of varieties with large fixed costs. This possibility of producing highly
mechanized crops provides additional profits for the farmer, which is further translated into
a higher demand for non-tradeable goods, and a larger expansion of the urban center.
To analyze the empirical implications of the model, I combine two sources of exogenous
variations that affected the export oriented production in the region. First, the construction
by the Military government of a large federal road system in Brazilian Amazon in the end of
the 1960s . The project consisted of a nearly 20,000 km highway to offer the first set of road
access between the Amazon and the rest of Brazil. Before the construction of the federal
roads, most of the transportation of goods was carried through waterways, and there was
very limited alternative routes in-between counties. The main goal of the federal government
was to connect State capitals in the region and with the federal Capital. Given this policy
design, I follow a growing literature evaluating the impact of transportation infrastructure
(Redding and Turner, 2014), and I compare counties in-between nodal cities that happened
to be closer to the transportation infrastructure with counties farther away 2.
Second, I explore the increase in commodity prices in the end of the 1990s, which was
directly related to variations in the Brazilian exchange rate and the increase in commodity
prices in 2007. Variations in the exchange rate were largely determined by the federal
government’s efforts to reduce inflation, which reached a level of 2000% in 1993. In 1994,
a fixed exchange rate pegged to the dollar was adopted, and, in February of 1999, the
Central Bank left this policy to adopt a inflation target one with floating exchange rate.
As a consequence, the Brazilian currency devaluated in 300% from 1999 to 2003. In 2007,
when the Brazilian exchange rate was back to lower levels, the price of commodities in the
international markets soared due to China’s economic growth. Combined, these shocks on
price sustained a decade long period of high profitability in the production of export oriented
commodities in the Amazon.
For the empirical section, first, I proceed with an analysis of the reduced form implications
of the model. I use decadal census data to analyze differential trends in population and
2See (Redding and Turner, 2014) for a review of the literature on the evaluation of the impact of trans-portation projects.
3
income between counties according to their access to federal roads. I show that there is no
statistical difference between counties close and far from roads before the 1970s in terms
of their population, and a limited one until the end of the 1980s. We observe, however, a
large divergence between counties close and far from roads in terms of their population and
income after the 1990s. Using annual information on commodity production and employment
in the urban and rural sector, I show that the expansion of economic activities in terms of
commodity production and employment follows closely the annual increase in commodity
prices. During this expansion, there is a widespread expansion in commodities with low
fixed costs of production (cattle), but only counties close to federal roads specialized in
the production of commodities with large fixed costs for capital (soybeans, corn and cotton).
Furthermore, I show that, in spite of a large expansion in the rural production of the Amazon,
we observe a limited expansion of employment in the rural sector, but a large one in activities
typically associated to urban centers (such as commerce, construction, and services).
In a second empirical section, I estimate the structural parameters of the model using
a version of the simulated method of moments. The estimated model has a good fit with
the data and is able to tie the relationship between roads and the key variables of the
reduced form analysis into an unique theoretical framework. Furthermore, in the absence of
a source of exogenous variation that provides comparable counties with different production
structures, the estimated model allows me to investigate what would have happened in the
absence of a capital intensive sector in the region. In the structural model counterfactual, I
find that almost all the job generation in the urban centers can be attributed to the presence
of a highly profitable activity with large fixed costs.
The results from this article provides three main contributions for the literature in trade
and development. First, this is one of the first studies to provide microeconomic evidence of
the theoretical mechanisms relating commodity trade to urbanization.3 Second, the results
from this article suggest that trade and the adoption of large-scale techniques for agricultural
production may be central elements to understand urbanization in developing countries
(Foster and Rosenzweig, 2007, 2004). Third, there is a recent and already large literature
evaluating the impact of transportation projects (Redding and Turner, 2014). However, to
the best of my knowledge, this study is the first to explore additional conditions that may
foster the realization of the benefits of such investments. In particular, I show that the lack
of good macro-economic conditions for trade prevented the urbanization of counties near the
3There are two recent articles on a similar vein. See Jedwab (2013) for evidence of the impact of commod-ity production in Ivory Coast and Ghana, and (Fajgelbaum and Redding, 2014) for an analysis of commoditytrade and structural transformation in Argentina.
4
transportation network.
The remainder of this article is organized as following. In the next section, I discuss the
evolution of the Brazilian Amazon economy and the main stylized facts related to population,
employment, and agricultural production in the region. In section 3, I develop a model to
guide the empirical analysis. In section 4, I study the reduced form implications of the
model. In section 5, I estimate the model structurally and I follow with the counterfactual
analysis. Finally, in section 6, I conclude the article.
2 The Brazilian Amazon Economy
2.1 Earlier Occupation and The National Plan of Integration (1969)
In 2010, around 10% of the Brazilian population lived in the Amazon region. For most of the
21th century, however, the region was largely unnocupied and sparcely populated. Before
the 1970s, most of the transportation in the region was carried through waterways, and there
was very limited land connection in-between cities (Pfaff et al., 2007). Using data organized
by the Instituto de Pesquisa Economica e Aplicada (IPEA) on population from all the census
available in Brazil, we can draw the population’s trend from the very beginning of the 21th
century and visualize the importance of the federal programs in the Amazon initiated during
the 1970s.
The dotted line in figure 1 shows the trend in the Amazon’s share of the Brazilian
Population. Before the 1960s, the Amazon region represented around 4% to 6% of the
Brazilian population across different census years. Furthermore, the solid line shows that,
before the 1960s, at most 3 million people lived in the Amazon region, which has more than
4 million km2 (roughly half the area of the United States). Interestingly, the population
data captures the period of economic expansion in the region around 1910, when there was
an intense economic expansion associated with the export of rubber for the automobilistic
industry in the United States. During this period, the region established several central
cities that would later on constitute the bases upon which the federal government designed
a transportation project for the region.
In the end of the 1960s, Brazil was going through a period of unprecedented economic
growth (with an average of more than 10% a year between 1968 and 1973). During these
years, the Military government was running a dictatorship and financed a series of projects
to integrate the Amazon region with the rest of the country. The central one was the
National Plan of Integration (1969). The goal of the federal government with this plan was
5
twofold: first, to occupy the Amazon region and reduce the likelihood of its occupation by
neighbor countries; second, to reduce the flux of migrants to the industrializing centers of
the southeast. The central piece of the plan consisted in the construction of a nearly 20,000
km highway system to connect State Capitals in the Amazon between themselves, with the
Federal Capital, and with other regions of Brazil 4.
Figure 4 clearly shows that the federal intervention during the 1970s was a watershed
for the occupation of the Amazon. From 1970 to 2010, the total population of the region
increased in roughly 500% (from 3 to 18 million people). While part of this population’s
expansion may be attributed to a demographic transition in the region, the large magnitude
of the expansion indicates that there was an intense in-migration to the Amazon. Further-
more, other regions in Brazil were also going through a demographic transition during the
period and, still, the Amazon’s share of the Brazilian population increased from 4% in 1960
to 10% by the end of 2010, when all the regions had already completed their demographic
transition.
2.2 Shocks on Commodity Price
After the 1970s, and until the end of the Military’s era in the middle of the 1980s, the federal
government made several attempts to promote the agricultural production in the Amazon.
Farmers who migrated to the region were able to acquire large plots of land for extremely
low prices 5. However, the poor quality of the soil and the climatic conditions prevented
them from producing staples that were typical in their native regions (such as coffee and
wheat). For most of the initial years of occupation after the 1970s, the wood and the mineral
industries remained as the main sources of income for the recently migrated population.
During the 1990s, the conditions for the establishment of a large agricultural industry
in the Amazon were met due to a series of technological improvements and macro-economic
changes. First, new seed varieties developed by the National Agropecuary Research Insitu-
tion (EMBRAPA) allowed the production of soybeans in low latitude areas (Vera-Diaz et al.,
2008). Second, mouth vaccines controlling disease enabled ranchers to attend the southeast
and the international requirements for meat consumption (Walker et al., 2009). Third, there
was an increase of roughly 200% in the price of export oriented commodities in terms of the
4See figures A1 and A2 in the appendix for the original maps of the plan.5There are several histories of colonization companies that actually gave huge plots of land for free in
order to establish an initial community in the region. Interestingly, the slogan of the government to attractpeople was “Terra sem gente para gente sem terra”, which translated becomes “land without people forpeople without land”.
6
local currency which made the production of these commodities a highly profitable activity
(Richards et al., 2012).
Variations in commodity prices faced by farmers in the region, which is a central source of
exogenous variation for this study, is directly related to variations in the Brazilian exchange
rate and the international price of commodities. This claim is largely supported by data.
The dotted line in figure 2 shows the evolution of the international soybean price given by the
World Bank time series multiplied by the exchange rate in Brazil, obtained from the Banco
Central Brasileiro (BACEN). The solid line presents the price offered by local retailers to
farmers in the county of Primavera do Leste (one of the main rural producing counties in
the region) measured by the Instituto Mato-Grossense de Economia Agropecuaria (IMEA).
In spite of the level differences between these prices, the variation follows each other closely.
Using annual data on agricultural production from Producao Agrıcola Municipal (PAM),
we observe that, between 1999 and 2004, when the price of commodities soared, there was a
large expansion in the production of export oriented commodities. Figure 4 shows that both
the total production of cattle and the sum of the land area producing soybeans, cotton and
corn follows the increase in commodity prices. Furthermore, in spite of occupying a similar
area in the beginning of the 1990s, we do not observe any effect on the production of all the
other 15 crops combined 6. This large expansion in agricultural production did not come
without environmental costs. Between 1999 and 2010, there was an intense deforestation in
the Amazon (Richards et al., 2012). Between 2001 and 2004, when the annual deforestation
rates were at its highest rates in the past 20 years, almost 100.000 km2 of the Amazon
forest was cleared for agricultural production (Morton et al., 2006). Note that specific
technological improvements may have provided important conditions for the expansion of
these commodities, but they are less likely to explain the timing of expansion in both the
cattle and in the soybeans production (besides the fact that they might be endogenous to the
profitability of production). Furthermore, aside the expansion in these commodities, several
other export oriented crops were affected by the exchange rate shock between 1999 and 2004
in other regions of Brazil. In the appendix, I show a more detailed analysis of the expansion
of export-oriented crops between these years.
In spite of a drastic expansion in the agricultural production of the region, annual ad-
ministrative data from the Registro Anual de Informacoes Sociais (RAIS) on all the formal
employment in the region indicates a larger expansion in the generation of jobs in sectors
6Other crops in figure 3 includes: manioc, wheat, rice, peanuts, bananas, potatoes, cocoa, coffee, sugar-cane, onion, beans, tobacco, orange, tomatoes, and black pepper.
7
typically associated with urban centers. The solid line in figure 4 combines commerce, con-
struction, and service sectors, which I define as the urban sector for the remainder of this
article. The dotted line presents the amount of jobs in the agropecuary, defined as the rural
sector. Between 1990 and 1998, employment in the urban and the rural sectors follow a
parallel trend. However, after 1998, we observe a divergence between their trends 7.
To summarize these stylized facts, I present in figure 5 all the relevant variables from this
section normalized according to their mean and deviation across the years. In 1998, before
the commodity price increase, most of the variables were in levels comparable to those in
the beginning of the 1990s. In the first five years following the devaluation of the currency
in 1999, all the variables achieved unprecedented levels that were sustained until the end of
2010.
2.3 Technological Characteristics of the Agricultural Production
Before moving to the next section, I present central characteristics of the agricultural pro-
duction in the Amazon which will be used to define the theoretical model.
First, as indicated in figure 3, a few rural activities dominated the dynamics of the
agricultural production in the Amazon after the 1990s: soybeans, corn, cotton, and cattle.
One important characteristic of production is the complementarity between soybeans, corn,
and cotton, but the substitutability between soybeans and cattle. While soybeans work as
the leading activity, farmers also produce corn in a second harvest season within the year to
generate additional profits 8. Furthermore, the same machines used to harvest soybeans can
be applied on the production of corn. Cotton is usually produced as a additional activity
and to diversify the risk and increase profits according to the expected prices.9 In figure
E1 in the appendix, I show a clear positive correlation between soybeans, corn and cotton
7The industrial production in the region outside the city of Manaus is very incipient. Manaus is anexception, because the federal government created a free tax zone in the city to give incentives for people tomigrate. As a capital of one of the States, this city was excluded from the sample for the following analysis.
8Harvesting twice within a year is typically called as double cropping. However, according to the largemajority of farmers interviewed in the region, the second season is much less productive and works at most asa limited complement to their production. As such, this technique seems not to be doubling the endowmentof land in practice. In general, farmers would have to spend resources covering the land with some typeof grass vegetation to protect the soil nutrients from the sun between the months of June until September,when the cultivating season restarts. They take this opportunity to use the land to cultivate something thatcan actually be sold and generate some additional profits given that labor and capital was already hired forthe year. In general, a full typical corn production is not adopted, but a less expensive variety of seed calledMilheto.
9The production of cotton, however, is very risk, which prevents farmers from producing cotton as aleading activity.
8
at the county level, but a negative one with cattle production. The data also shows that a
few counties have no production of corn and cotton in the presence of a substantial soybean
production. There are several counties, however, with a small production of soybeans but a
large production of cattle, which is consistent with a substitutability between these activities.
To simplify the following investigation, I combine the land area use for soybeans, cotton,
and corn into one.
Second, there is a key difference between the production of cattle and soybeans: the latter
requires large amount of investments in machines and infrastructure such as harvesters and
seeders, which are generally imported from abroad. Also, the production of soybeans requires
large amounts of imported fertilizers and limestone to neutralize the acidity of the soil and
enrich it with nutrients for the crops in the beginning of the season. Cattle production,
however, requires much lower investments in machines, and a very limited preparation of the
soil. In fact, in table D9 in the appendix, I use data from the Agricultural Census in 2007 to
show that there is a stronger correlation between large tractors and soybeans with respect
to cattle, but the inverse is true for small tractors (where large tractors are defined as those
with more than 100 horsepower). The production of cattle, however, generally precedes
the production of soybeans after clearing the land. Therefore, one could argue that cattle
production takes some extra costs for clearing the land. The available data does not allow a
complete assessment of all the fixed costs involved in the production of soybeans and cattle.
Still, it seems plausible to assume that the fixed costs of producing soybeans are larger than
the ones related to cattle production. In the structural estimate of the model, where I allow
these fixed costs to be arbitrarily different, I find that the production of soybeans have a
larger fixed costs than cattle ranching.
3 A Theory for the Amazon Economy
The goal of developing a theoretical model for the region is twofold: first, to study the
relationship between the endogenous variables and the sources of exogenous variations in
a set of regressions; second, to estimate the structural parameters of the model to explore
counterfactuals that would not be feasible under a reduced form approach. I take advantage
of the specific characteristics of the Amazon economy to impose assumptions consistent with
the data that will keep it tractable for empirical estimation. The main dynamic force in the
model is generated by variations in the level of commodity prices common to all counties, and
a spatial variation that is given by distance to federal roads. These two elements constitute
9
the main sources of exogenous variation for the empirical analysis in the following sections.
Here, I present the main equations of the model. In the appendix, I provide its full derivation.
3.1 The model
The production is divided into a rural and an urban sector. In the rural one, each county i
has one immobile representative farmer who decides between producing two export oriented
activities: soybeans (S) or cattle (C).10 Soybeans represent the modern large scale produc-
tion system that also includes corn and cotton. Because other crops are less important for
the economic dynamics of the region and to the patterns of land use (as shown in figure 3), I
abstract from them to simplify the model. The land use decisions are based on the Span-of-
Control model from Lucas (1978). The farmer decides what to produce in each piece of land
` according to prices, technology, the land quality z(`), and the fixed cost associated with
each commodity (fS and fC). Each county has a measure of land Ti such that ` ∈ (0, Ti].
The urban sector produces a non-tradeable good that only takes labor as input. This sec-
tor represent restaurants, education, housing, and other nontradeable activities. Workers in
both the rural and the urban sector can migrate between activities and between cities as in
typical open city models (Roback, 1982). This is consistent with patterns of migration in
Brazil, where 34.5% of the population lives in a city different from the one where they were
borned (Census, 2010).
Consumers
Each county i is home to Ni mobile workers and a representative immobile farmer. Both
of them have cobb-douglas preferences for a manufacturing good (cM) produced elsewhere
and non-tradeable goods (cU) produced in the urban center of each county. I assume that
soybeans and cattle are not consumed directly by these counties and that, for final consump-
tion, they have to be processed elsewhere. The utility function is given by:
U = δcαMM cαUU (1)
where I set δ = 1ααMM α
αUU
for algebraic convenience and without loss of generality.
The price faced by consumers depend on the distance a county is from a federal highway.
10The results of the model are not dependent on this assumption. I could assume, instead, that thereis an arbitrary number of farmers, each owing a share of the county area. As long as the preferences arehomothetic, the consumption of each farmer can be aggregated into one single farmer. What is importantfor the results, however, is that farmers are immobile.
10
Manufacturing goods have to pay an iceberg cost according to an iceberg decay parameter γ
to reach the county. Therefore, manufacturing goods are more expensive farther away from
roads. I assume the transportation costs to follow an exponential decay.
pMi = p∗MeγDroadsi
Where p∗M is given exogenously and Droadi represents the distance a county i is from a
federal road. The price of non-tradables, pUi, is given endogenously by supply and demand
within the city.
Workers may leave the county and migrate to other regions of Brazil if a minimum utility
is not achieved. They must be indifferent between living in county i, and an exogenous
outside option providing utility V . Manipulating the indirect utility function provides the
following non-migration condition.
V =wi
eαMDroadsiPi(2)
Where Pi = (p∗αM
M pαUUi ).
Rural Sector Production
Farmers may produce two export oriented activities: cattle or soybeans. The latter
represents the modern activity with large fixed costs, and the former the traditional one with
low capital requirements. The technologies for each are given by the following equations:
yS(`) = aSz(`)lφ(`)
yC(`) = aCz(`)lφ(`)
(3)
yS(`) and yC(`) are the outputs for producing in the unit of land `, l(`) is the amount of
labor employed, aS and aC are parameters related to the productivity of each activity, and
z(`) represents the distribution of land quality within the county - which could be associated
with differences in soil quality, slope or distance to the urban center. For simplicity, I assume
that both technologies have the same labor intensity factor φ. 11 Note that we can think
of these technologies as constant returns to scale where land intensity is given by 1− φ and
land employed is equal to 1. The production of each activity requires a fixed cost fS and fC
11Note that, even though I assume the same φ, I do allow different technologies to have different demandfor labor given the use of land. Because I am holding the land in piece ` to be 1, the parameter aS generatesdifferences in the demand for labor for each given piece of land.
11
such that fS > fC .
The prices faced by farmers, which is a central for the dynamics of the model, take the
following functional form:
pCi = P ∗Ce−γCDroadsi
pSi = P ∗Se−γSDroadsi
(4)
P ∗C and P ∗S are the international price adjusted by the exchange rate for soybeans and
cattle respectively. Variations in P ∗C and P ∗S generate level variations that are common to
all counties in the Amazon region. e−γDroadi represents the spatial variation in the level of
prices and does not change over time. These two elements combined capture the two sources
of exogenous variations that I explore in the empirical section of the article to evaluate the
effect on the endogenous variables of the model.
The profit function for each piece of land ` is given by:
π(`) =
maxl(`){pSe−γSDroadiaSz(`)l
φ(`) − wil(`) − fS} if yS(`) > 0
maxl(`){pCe−γCDroadiaCz(`)l
φ(`) − wil(`) − fC} if yC(`) > 0
0 if yS(`) ≤ 0 and yC(`) ≤ 0
The landowner maximizes the profit in each piece of land according to the potential
profits of each activity and the possibility of not producing anything:
π(`) = max{πS(`), πC(`), 0} (5)
This framework provides a clear economic motivation for decisions on land clearing and
the consequences for deforestation. As profits of agricultural goods increase, the opportunity
cost of the land increases and farmers clear the forest to produce agricultural goods. In the
model, the fixed cost of setting up agricultural activities prevent farmers from clearing all
the land of the county for production.
Urban Sector Production
The urban sector produces non-tradeable goods according to a constant return to scale
technology with no land requirement:
12
yUi = aU lUi (6)
Where yUi is the total output, aU is the productivity of the sector and lUi is the amount
of labor employed.
Competitive Equilibrium
Given a vector of prices p = (p∗S, p∗C , p
∗M), a distance from roads Droadi, an exogenously
given indirect utility function V , and a land endowment Ti, the competitive equilibrium in
each county i is characterized by a population Ni, labor demands {l`}`∈(0,Ti] and LUi, and
patterns of land use such that:
1. (Non-migration Condition) Workers maximize utility
Vi ≥ V (7)
2. (Rural Sector Production) Landowners maximize profits in each plot of land `
π(`)i = max{πS(`)i, πC(`)i, 0} (8)
3. (Urban Sector Production) Firms maximize profits
πU ≤ 0 (9)
4. Labor market clears
∫`∈TSi
lS(`)d`+∫`∈TCi
lC(`)d`+ LUi = Ni (10)
5. Trade is balanced
3.2 Comparative Statics
To study the comparative statics, I define a set of parameters that makes the partition of
land in the model consistent with some basic facts of the data. In the structural estimates,
however, I allow the parameters to be arbitrarily different.
13
Proposition 1 (Partition of Land) Define zSCi as the plot ` where the farmer would be
indifferent between producing cattle and soybeans, and zCFi for the plot ` where the farmer
would be indifferent between producing cattle and clearing the forest. If the conditions pS >
pC, fS > fC, and zCFi < zSCi are satisfied, then the area of the county can be partitioned
in three where soybeans is produced in ` such that zCFi < z(`) < zCSi , cattle is produced in
zCFi < z(`) < zCSi , and forest is left uncleared for 0 < z(`) < zCFi .
Proof In appendix.
The parameters required for the proposition to hold are plausible: soybeans must be rel-
atively more profitable than cattle ranching (p∗SaS > p∗CaC) and the fixed costs of producing
soybeans must be larger fS > fC such that there is a tradeoff between the higher gross profits
derived from soybeans and its fixed costs. Further, I need cattle to be sufficiently profitable
with respect to soybeans, otherwise, we could still have full specialization of the county on
soybeans (see appendix for details on other possibilities of partition of land). Condition
zCFi < zSCi is sufficient for the county to have a positive production of soybeans and cattle.
The partition of land into zCSi and zCFi is generated by a combination of differences in
fixed costs and the complementarity between land quality and prices. One can show that
the cross derivative between prices with respect to land quality is positive. Therefore, as
we go from a plot with low quality to higher quality ones, this movement has a larger effect
on the gross profit of soybeans than on the one for cattle. However, the difference in fixed
costs prevents farmers from producing soybeans in every plot. There will be a range of land
quality between zCFi and zCSi where farmers prefer to produce cattle in spite of the smaller
gross profits, because profits net of fixed costs are larger for cattle production (given that
fS > fC).
Following the parameters in proposition 1, I first show a set of results for the model
that should hold given the level of prices p∗C and p∗S. Second, I study how variations in
p∗C and p∗S affect the production of counties given the distance from roads eγDroadsi . Given
the general equilibrium effects within each county, the model does not allow for closed form
partial derivatives. In the appendix, I show a set of figures with simulations where I study
the comparative statics of the model.
Variation in Distance to Roads given International Price
For a given international price level p∗C and p∗S, decreasing distance to roads increase the
land area dedicated to soybeans, the total income of the county, the population, and the
14
employment in the urban center. However, the effect is ambiguous for the land used for
pasture and may follow a U-shape depending on the parameters of the model.
The intuition for the possibility of a U-shape relationship between cattle production and
distance to roads is the following. When we increase distance to roads, the price faced by
farmers in the county for commodities is smaller due to the transportation costs. At the
marginal land producing soybeans (zi), cattle becomes more profitable because of its lower
fixed costs, and there is a reduction in the area producing soybeans but an expansion in the
one producing cattle. However, at the marginal land where the farmer is indifferent between
clearing the forest and producing cattle (zi), an increase in distance to roads makes the
production of cattle unprofitable, decreasing its use of land area. Therefore, at one margin,
there is an increase in the amount of land producing cattle while, at the other, there is a
decrease. The effect of distance on the production of cattle is thus ambiguous and depends
on the magnitude of the effect on each margin
Variation in International Prices given Distance to Roads
When international prices p∗C and p∗S increase, there is a positive impact on the land area
used to soybean production, the total income in the city, the population, and the generation
of jobs in the urban centers. The effect of the price increase is magnified by the access
to federal highways, i.e., counties closer to roads have an even larger expansion on these
endogenous variables. The impact of the increase in commodity prices on cattle production,
on the other hand, is heterogeneous and depends on the initial production before the price
shock.
Here, the main intuition behind the heterogeneity is that cattle ranching precedes the
production of soybeans. If a county has an already large production of cattle before the
positive shock on commodity prices, there will be an expansion of soybeans over land used for
pasture, with limited expansion of pasture over forest areas (which are the remaining areas of
the county where the productivity of land is already too low to justify even the production of
a low fixed cost activity). However, in counties that have a small production of both soybeans
and cattle production in an initial period before the shock, there will be first an expansion of
the activity with lower fixed costs over forest areas, i.e., cattle. Note that the model is able
to rationalize one of the main stylized facts discussed in the literature about deforestation
in the Amazon (Barona et al., 2010): a weak correlation between soybean production and
deforestation, but a strong correlation between cattle ranching and deforestation.
15
4 Reduced Form Evidence and the Dataset
4.1 The Dataset
For the empirical investigation, I combined information from a series of different sources to
organize a dataset at the county level with economic production variables and geographic
characteristics related to agricultural productivity. A thorough description of the source of
information for each variable is given in the appendix. Figure 6 presents the main structure
of the dataset, where I present the county boundaries in 2010, the federal highway system
constructed, and the project lines according to the documentation in the National Plan of
Integration of 1969.
The information available allowed me to construct two panels at the county level. One
that goes back to the 1950s with decadal information on population and income. And another
that starts in 1990, which includes annual information on commodity production and jobs
in the formal sector. Because several counties were created between 1970 and 1990, I used
information available from the Brazilian census bureau (IBGE) on the date of creation of
each county to improve consistency of the boundaries over time 12.
I included in the dataset several controls that are potential determinants of agricultural
productivity. First, I brought information on rainfall and temperature in September obtained
from the WorldClim (Global Climate Data). This month is a crucial one for the beginning
of the harvest season, when the rainfall patterns determine the decisions for seeding the plot.
Furthermore, I added information on corn and soybeans suitability from the FAO Global
Agro-Ecological Zones, which is calculated based on several geographic characteristics of the
region and has been used in recent articles (Bustos et al., 2013) 13. Finally, I also calculated
the average slope of the county using information on altitude provided by the NASA’a Shuttle
Radar Topography.
The central explanatory variable in the analysis is distance of the centroid of each county
to a federal highway constructed during the 1970s. One caveat about this measure is that
I am implicitly assuming that being farther away from a highway does not make a county
closer to another transportation system. To avoid this problem, I excluded all the counties
12I used the information provided by the Instituto Brasileiro de Geografia e Estatıstica to create an algo-rithm that aggregates information of each year according to the set of counties that were already created inthe initial year of the dataset (i.e., 1950 and 1990). The dataset provided by IBGE contains information onthe date of creation of each city as well as the cities from which each city was derived.
13The FAO Global Agro-Ecological Zones data provide suitability for different levels of input use. Becauseof the large-scale pattern of production in the region, with several farmers producing in farms with over 500ha, I included information on high input.
16
further than 300 km from a highway from the analysis (around 1% of the sample) 14. An-
other caveat is that federal highways in the Amazon have different qualities. Actually, the
quality on federal highways is far more heterogeneous than the quality off federal highways.
Furthermore, in the sample, most of the roads were paved several years after the implemen-
tation of the National Plan of Integration. Therefore, the quality of the road is more likely
to be correlated with further economic expansion of the counties than its initial condition.
In spite of that, to address this concern, I control in the empirical section for region fixed
effects, which makes counties under comparison more likely to face the same quality of the
transportation network.
4.2 Empirical Strategy
For an initial evaluation of the differential impact of the commodity price shock on the
counties of the region according to their access to the federal highway, I use the following
specification:
Yit − Yi(t−1) = θClosei +Xiβ + αj + εi (11)
Where the subscript t denotes a year, j a region fixed effect, and i a county. Closei is a
dummy equal to 1 if the county is closer than 100 km from a federal highway, Xi is a set of
control variables, and εi is the unobserved component. θ is the parameter of interest.
Equation 11 is estimated for different time periods t, for example, t = 1991 and (t− 1) =
1970. In this case, I estimate the impact of the federal highways on the economic activities
before the shock in commodity prices. For t = 2010 and (t−1) = 1991, I estimate the impact
of being close to the federal highways after the commodity price shock. For the dependent
variables with annual frequency, I am able to explore with more precision the timing of the
price shock. For these variables, I set a regression for t = 1999 and (t − 1) = 1994, and
another for t = 2004 and (t − 1) = 1999. This time period structure, with periods before
and after the price shock, leads to the final specification to estimate θ:
After Price Shock︷ ︸︸ ︷(Yit − Yi(t−1))−
Before Price Shock︷ ︸︸ ︷(Yi(t−1) − Yi(t−2)) = θClosei +Xiβ + αj + εi (12)
To identify the parameter θ on equation 12, I need to assume the differential impact
14Another problem is that there are only counties in the far northwest of the Amazon that are fartherthan 300 km from a highway network. To some extent, these counties are in very remote areas, with almostno meaningful economic production, and provide very little variation for the analysis of this study.
17
of prices on counties that are close to highways to be uncorrelated with unobserved shocks
that coincides with both the timing of the price shock, and the location of counties. Note
that year specific technological improvements do not violate the identification assumption, as
long as they are sufficiently accessible for farmers who are both close and far from highways.
However, the assumption might be violated if being close to a road is correlated with the
suitability for production. For example, if the route of highways were constructed to avoid
steep areas, counties close to them would be flatter and more suitable for a mechanized
production. Once we have the price shock, these counties would potentially increase more
their agricultural output with respect to counties farther from the highways because of their
better geographic characteristics for production.
An analysis of the control variables can shed light on the plausibility of the identification
assumption. Table 1 presents the geographic variables included in Xi used as controls in
the estimation of equations 11 and 12. In columns 2-3 and 5-6, I present the relationship
between the control variables and distance to roads in elasticity to facilitate the interpretation
and the comparison of the coefficients. In general, panel A of table 1 presents a weak
correlation between distance to roads and the geographic variables in both statistical and
economic terms. The area of the county, however, is an exception and has a large and
significant correlation with distance to federal highways. In part, this is due to the mechanical
correlation between distance to federal highways and State capitals generated by the design
of project. Counties closer to the State Capitals are generally smaller, and counties close
to the federal highway system ended up being those closer to State Capitals. In spite of
this, the combination of a weak correlation with geographic variables and a strong one with
distance to State Capitals reinforces the fact that the main determinant of roads was the
government‘s goal of connecting State Capitals. Given this framework, in all the estimates
of equation 11 and 12, I control for a cubic polynomial on distance to State Capitals.
Another evidence that supports the validity of the identification assumption comes from
an investigation of the population before the construction of the federal highways. In table
1 panel B, while the full sample in column 2 presents a moderate and statistically significant
correlation between distance to federal highways and population levels before the implemen-
tation of the transportation project, once we drop State capitals from the sample in column
3, the magnitude of the correlation reduces to almost zero in 1950 and by half in 1960. This
provides evidence that there was limited connection in-between counties, and that the route
of the federal roads were not constructed to benefit specific counties with larger economic
production. For the following analysis, I drop State Capitals from the sample and compare
18
only counties in-between.
Several recent articles on the evaluation of transportation projects use the least-cost
spanning tree network as an instrument for actual distance to federal roads (Faber, 2014).
The main goal of this instrument is to obtain a source of variation that is uncorrelated with
the demand for transportation, which could lead to a simultaneity bias with the expansion
in economic activities. In the present context, however, it is unlikely that the federal govern-
ment planned the route of the roads according to an economic boom that occurred around
30 years later. Furthermore, this instrument would mechanically increase the correlation be-
tween distance to roads and geographic characteristics. Given that agricultural production
is at the core of the Amazon’s economy, this would raise concerns about omitted variable
bias.
Another option of instrument would be to use the project lines in the documentation
of the plan (Baum-Snow, 2007). Table 1 panel C presents a strong correlation between
actually constructed highways and distance to project lines, which leads to a strong first
stage. However, investigations of the correlation between project lines and the covariates in
table 1 does not provide evidence that the instrument is less correlated with potential omitted
variables. Actually, when highways are instrumented by the project lines, the magnitude of
the coefficients are very similar, but statistically weaker, which led me to opt for the ordinary
least square approach.
4.3 Results
In this section, I present the results from estimates of equation 11 and 12 using the main
endogenous variables from the theoretical model as dependent variables.
Population and Income (1950-2010)
As initial evidence of the impact of the federal highway system in the Brazilian Amazon
and the commodity price shock, figure 7 presents the difference between counties close and far
from federal roads on population and income for each census year. From 1970 to 1980, in the
first 10 years after the beginning of the implementation of the National Plan of Integration
in the region, there is a divergence in trends between counties close and far from federal
roads. This divergence process slows down during the 1980s. In the following 20 years
after the 1990s, however, we observe an intense new process of divergence between counties,
which corresponds to the period when the economic production of commodities for external
markets expanded in the region.
19
Table 2 presents the results from the specification 11 for the time lag that captures the
first 20 years after the implementation of the project, and another that captures the following
divergence after the 1990s. Column 1 shows a large impact of the construction of roads on
population. Given the baseline level (reported in the table) the magnitude of the coefficient
reinforces the fact that the impact is large and that the initial level of occupation in the
Amazon was low. When we control for a series of geographic variables and initial conditions
in column 2, the coefficient remains relatively stable. Note that a F-test of the joint statistical
significance rejects the null-hypothesis of no joint effect of the control variables at the 1%
level (p-values are reported in the table). This is a systematic pattern across most of the
specifications in this reduced form section, which shows that the controls have a strong
predictive power on the dependent variable. In column 3, where I present the results for the
period between 2010 and 1991, we can see a strong divergence process. The impact of being
close to roads is much larger during this period. In column 5, I estimate equation 12 to assess
the impact of commodity prices on the trends between 1991-1970 and 2010-1991. The effect
of being close to roads confirms an intensification of the divergence after the 1990s. Also,
note that the controls for geographic characteristics have a limited effect on the magnitude of
the coefficients across all the specifications, which reinforces the fact that distance to roads
is orthogonal to characteristics related to agricultural productivity.
The divergence between counties close and far from roads after the 1990s is clearer when
we look at income, which is presented in panel B of table 2. There is a very limited divergence
between 1991 and 1970. However, a strong one after the 1990s, when the export oriented
agricultural industry emerged in the Amazon. In general, all the qualitative conclusions from
population can be extended for the analysis of income.15 Coefficients are large in magnitude,
stable across specifications, and they suggest a large divergence after the 1990s, more than 20
years after the construction of the federal road system. However, the fact that total income
did not follow closely the population expansion between 1970 and 1990 is puzzling. In part,
this could be due to the fact that in the first 20 years after the construction of roads, there
was still a large proportion of the population living from a subsistence production, which is
poorly captured by a monetary measure such as income. Unfortunately, good measures of
subsistence production is not available for earlier years. Still, for the scope of this article,
15Here, I use total income because it is more closely related to total population than income per capita.Furthermore, the model does not provide any theoretical prediction on the impact on income per capita.Mainly, because of the high migration rates across counties, the gap in income per capita should be small.In fact, in the next section, I show evidence that wages are very similar across different counties independentof their distance to federal highways.
20
it seems sufficient to document that using both population and income variables to analyze
trends in the economy of the Amazon leads to the same conclusion: that an additional
shock occurred after the 1990s that intensified a divergence in economic production between
counties close and far from federal roads.
Commodity Production (1990-2010)
The data on population and income does not allow me to explore with more precision the
timing of the increase in commodity prices. Here, I use annual data on commodity production
to evaluate the divergence in trends between the periods of 1999-1994 and 2004-1999.
As an initial assessment of the differential trends, figure 8 shows the difference in the
average production of commodities between counties close and far from federal highways.
The solid line in the figure presents a clear divergence between 1999 and 2004, when the
commodity prices reached a historical peak. In spite of the aggregate cattle production
following the trend in soybeans during this period (see figure 3), we do not observe any
differential trend here. According to the model, this could be attributed to the highly
heterogeneous impact of the price shock on cattle production. To investigate this point,
figure 9 shows the trends in cattle production in counties that reached 2010 being large
producers of soybeans (above the 95th percentile), medium producers (between the 95th and
the 90th percentile), and the rest of the counties in the Amazon (below the 90th percentile).
Between 1994 and 1999, trends in the production of cattle follow a relatively similar path
across different counties. However, after 1999, counties that reached 2010 as large soybean
producers have a clear decrease in their production of cattle, while counties with a small
production have an expansion in their cattle production of more than 60%. Interestingly,
the rank also changed during this period. Counties that became large producers of soybeans
in 2010 went from being large to small producers of cattle before and after the commodity
price shock. Counties that reached 2010 as small producers of cattle, however, went from
being small to large producers of cattle.
Table 3 shows the estimates of equation 11 and 12 using commodity production as the
dependent variables. As suggested in figure 8, there is a large impact of prices in the
production of soybeans between 2004 and 1999 as shown in column 3 and 4 of panel A.
Between 1994 and 1999, however, the coefficients suggest a process of convergence between
the production of counties close and far from highways. Taken together, the estimated effect
for the period after and before the shock suggest a divergence in the soybean production of
counties close and far from roads that coincides with the period when the price of export
oriented commodities were increasing. The difference in trends, presented in columns 5 and
21
6, confirms a divergence process in soybean production. In general, the statistical properties
of the coefficients are good: they are statistically significant at the 5 % level, and they are
not affected by the inclusion of controls. The magnitude of the coefficient is also substantial
when we look at counties far from roads in the initial period (given by the baseline).
When we look at all the 15 remaining crops in panel B, the results are much smaller
in economic terms, and statistically not different from zero. This suggests that during
this period, there was a very limited variation in the production of other crops besides
soybeans and cattle. This result support the modeling decision to abstract from these crops.
The results for cattle production, which are presented in panel C, are unstable. 16. As
discussed above, this can be attributed to the highly heterogeneous impact of prices on
cattle production.
Employment in the Urban and Rural sectors (1990-2010)
So far, the reduced form section has shown results that are consistent with the model:
the shock on commodity prices has a positive impact on production which is magnified by
the proximity to federal roads. In other words, when conditions for exporting commodities
improved, counties closer to the transportation system were disproportionately benefited by
the facility to export their products. Here, I evaluate how much of this effect is translated
into an expansion of the urban centers in the Amazon.
As in the previous analysis, I begin by showing the difference between counties close
and far from federal highways over time in figure 10. Again, the timing of expansion in
the generation of jobs follows the increase in commodity prices in the region. Figure 1A
indicates that the difference between counties close and far from federal highways rose after
1999. However, when we look at both of them in the same axis in panel B, we can clearly
see that the magnitude of the expansion in the urban sector was much larger than in the
rural one. This result supports the fact that the commodity expansion had an effect on the
generation of jobs in the urban centers of the Amazon. In spite of the agricultural production
being at the core of the economic dynamics of the region, the impact of the commodity price
shock had a far larger impact on the generation of jobs out of the rural sector.
Table 4 confirms the evidence from figure 10. Columns 1 and 2 present the results for the
period before the price shock. Between 1994 and 1999, there was a very limited expansion
of jobs in both the urban and the rural sector. On the other hand, the estimated effect for
16Unfortunately, I can not proceed with a similar analysis as the one in table 3 because a dummy equal toone if the county ended up producing soybeans is highly endogenous. When I proceeded with such empiricalexercise, the coefficients were highly unstable across specifications.
22
the period between 2004 and 1999 is much larger, with a greater magnitude for the urban
sector. The estimated effect of the commodity prices on the change in trends in column 5
and 6 show the same patterns.
4.4 Robustness
In this section, I proceed with additional exercises to address potential concerns on the
interpretation of the results.
Non-linear effects
To increase statistical power and simplify the interpretation of the results in the previous
section, I adopted the simpler approach of defining a dummy variable that indicates whether a
county is close or far from the federal highway system. Here, I show that non-parametric esti-
mates confirm the reduced form effect. Figure 11 shows non-parametric regressions adjusted
by a cubic polynomial on the distance from State Capital using the procedure suggested in
Robinson (1988) 17. The non-parametric estimates reveal that part of the results could be
driven by a set of counties that are very close to roads. In fact, some counties that are just
on the federal highways became important service centers for the region. Next, I study the
robustness of the results when I drop these counties from the sample.
Dropping counties specialized in the service sector
The theoretical model in this article does not account for the possibility of some counties
becoming service centers for the region. In part, most of the public institutions that provides
documentation and environmental permits for production are located in the State Capitals
of the region, which are already excluded from the sample. However, some counties along
federal highways became centers where farmers obtain services such as machine repairs and
bank loans. This is potentially an important issue for the interpretation of the results through
the lens of the proposed theoretical model. To deal with this concern, I present in table 5
the results from the main endogenous variables in the model dropping every county with
the centroid closer than 25 km from a federal highway. Because I lose around 20% of the
sample, I estimate the most parsimonious specifications where I control for a polynomial of
17The procedure resembles a Frisch-Waugh-Lovell theorem. First, I non-parametrically regressed thedependent variables and the controls against distance to roads and saved the residuals. Then, I regressed theresiduals of the dependent variables against the residuals of the controls ones to estimate a set of parametriccoefficients, as in the Frisch-Waugh-Lovell theorem. Finally, I run the non-parametric regression of thedependent variables subtracted from the predictions of the control variables using the estimated parameters.
23
distance to State Capitals. Table 5 panel A shows the results from the baseline estimates,
and in panel B the results when we drop counties close to the federal highways. In spite
of excluding 20% of the sample, which are counties that according to the non-parametric
regression could affect the results, the point-estimate results are statistically the same and
the qualitative interpretation is largely unaffected. While some counties may have become
important centers for the Amazon, this does not seem to be a first order issue driving the
results.
5 Structural Estimates
18 What would have happened to urbanization in the Brazilian Amazon in the absence of a
capital intensive crop? To answer this question with a quasi-experimental design, one would
need a source of exogenous variation to generate comparable counties with and without the
production of soybeans. This is clearly unfeasible in the present setting, because soybean
production is highly endogenous to suitability conditions, access to transportation, and com-
modity prices. However, under stronger assumptions, a structural estimate of the model can
provide us with counterfactuals to answer this question. Furthermore, a good fit of the
estimated model with the data would reinforce the plausibility of the mechanisms explored
in this article. In this section, first, I discuss the variations in the data that identify the
structural parameters of the model. Then, I present the estimation procedure, the results
from the estimation, and the fit of the model with a series of stylized facts from the data.
Finally, with the estimated parameters, I proceed with counterfactuals of the model.
5.1 Identification of the Parameters
Unfortunately, not all the parameters of the model are separately identified. First, pS and
aS, as well as pC and aC , are always combined in the equations defining the equilibrium
of the model. Therefore, I assume aS and aC to be equal to one. Also, I am not able to
disentangle a level effect of the productivity parameters from the level of the fixed costs.
In other words, the parameters of the model related to the productivity of soy and cattle
could all be multiplied by a factor, and, still, I would be able to fit the data as long as I
increase the level of the fixed costs. As such, I focused on the estimation of the difference
between the productivity of soybeans with respect to cattle, and I assume pC = fC = 10
18This is a very preliminary exercise. Comments are very welcomed.
24
and γC = 0.0001. The same argument follows for the estimation of the parameters of z(`).
I am not able to estimate both parameters from the extreme value distribution. Therefore,
I assume T , the parameter governing the location of z(`), to be equal to 1 19.
Also, with respect to the parameters related to consumption, aU and pM appear together
with the indirect utility function in the non-migration condition:
wi = (V eγαMDroadipαMM
aαUU)
1αM
There are no other equations defining the equilibrium which allows V , aU and pM to be
separately identified. Therefore, I assume aU and pM to be equal to one.
Even though I am not able to estimate many parameters of the model, I can still explore
a series of counterfactuals with the ones I can identify. The following vector of 8 exogenous
parameters in the model are estimated:
Σ = (θ, φ, pS, fS, γS, γM , V, αM)
Below, I discuss the identification of each of these parameters:
• θ: this parameter governs the dispersion of the extreme value distribution. The move-
ment of zSC ,zSF , and zCF (i.e., the marginal land where the representative farmer is
indifferent between soybeans versus cattle, soybeans versus forest, and cattle versus
forest) identifies this parameter.
• pS: the price of soybeans provides information on the level of production, land, and
labor employment. The difference between the employment in labor and land with
respect to cattle generates the differences in the demand for land and labor we observe
in the data.
• γS and γM : the parameters on the transportation costs generate differences in produc-
tion between counties close and far from highways. In other words, if there were no
transportation costs, the model would predict all the counties to have the same level
of production.
• fS: the absence of production in the model where soybeans have a better price than
cattle is only justified by the impossibility of obtaining positive profits in areas with
19The identification of the scale and the dispersion parameters of the Frechet is a common issue in theTrade literature. See (Eaton and Kortum, 2002).
25
lower land quality, which is determined by the size of the fixed costs. Furthermore,
note that if fS is sufficiently low, all the counties would specialize on the production of
soybeans (as long as pS > pC). The absence of a widespread production of soybeans,
therefore, also provides variation to infer fS from the data.
• φ: the parameter on labor intensity factor will be associated with the ratio between
labor to land. For low levels of φ, production would require a small ratio between land
and labor. Therefore, the amount of labor and land in the county together provides
the identification of φ.
• V : this parameter captures several pieces of the migration condition. Here, the level
of wages help me identify this parameter.
5.2 Estimation procedure
The method used to estimate the parameters of the model resembles the simulated method
of moments. First, I use the actual data for distance to roads and the area of the county to
generate a simulated dataset using the theoretical model. Then, I calculate a set of moments
from this artificial dataset with a given set of parameters, and I compare them to the actual
moments in the data. Finally, an algorithm procedure follows with several comparisons
between the moments generated by the model and the ones from the data to minimize their
difference.
Note that, in each interaction of the model, I have to calculate the general equilibrium
in each county, which makes this process computationally intensive. Therefore, to simplify
the estimation, I do not assume a specific structure for the noise generated by the data
as in a typical simulated methods of moments estimation - which are generally modeled
as productivity shocks or idiosyncratic utility preferences for living in a specific county20. Instead, I focus on a parsimonious number of moments that are not dependent on the
structure of noise in the data. More specifically, the following 10 moments are chosen:
Mxavg(Σ) = yx − yxs (Σ)
Mxβ (Σ) = βx − βxs (Σ)
(13)
20This can also be interpreted as a calibration where I use actual data to infer the moments from themodel.
26
Where x ∈ {La, Lu, TS, TC , w}, β is the coefficient of an ordinary least square regression
with a constant of the dependent variable x on distance to roads Droads, and s stands for
the moments coming from the simulated dataset. There are two sets of moments. First, the
average value of the dependent variables, which provides the identification for the parameters
determining the level of production in the region. Note that, by focusing on the average of
the endogenous variables of the model instead of their variance, I am allowing the actual
moment in the data to be matched by a deterministic model that does not incorporate the
noise in the data. The second set of moments are the coefficients from an ordinary least
square coefficients. They provide the variation to estimate the transportation costs in the
model. By using the ordinary least square coefficients as a set of moments, I attempt to
unify into a single theoretical framework the results that would be obtained by running 5
independent regressions on the relationship between the dependent variables and distance
to federal highways 21.
Defining M(Σ) = [Mxavg(Σ) Mx
β (Σ)] as the vector of moments. Σ0 as the true values of
Σ. The estimation is based on the condition that
E[M(Σ0)] = 0
The minimization algorithm searches for Σ to achieve:
Σ = arg minΣ{M(Σ)WM(Σ)′}
Where W is a weighting matrix. I choose the weighting matrix that equalize the order
of magnitude of the different moments used for the estimation.
For the reduced form section, I used data on formal employment in the Amazon region,
which is available on annual basis and allowed me to explore the timing of expansion in the
production of commodities. However, while using formal employment may provide the same
qualitative results as using informal one (as long as they are proportional to each other in
every year), for the estimation of the structural parameters of the model this would be a very
concerning limitation. For example, we would understate the value of φ, which is related to
the proportion between labor and land. To avoid this problem, I estimated the structural
model on data on land use available from the Agricultural Census of 2007, and the whole
21Note that, because I use the actual data on distance to roads for the simulated moments, using theordinary least square coefficients would provide the same results as using the covariance. However, I preferredto use the information on the coefficients as they provide a more intuitive way of relating the structural modelestimates to the reduced form one.
27
universe of employment (formal and informal) provided by the Census of 2010.22
5.3 Results
Table 6 presents the results for the fit of the model. In spite of a very parsimonious number
of parameters (8 parameters), the model is able to fit the data well. Panel A indicates a
very good fit of the simulated dataset with the actual data in terms of the average in the
endogenous variables of the model La, Lu, TS, TC , w. Furthermore, panel B shows that the
model is also able to reproduce broadly the qualitative patterns of the relationship between
roads and the endogenous variables. In particular, the coefficient for wages, soybeans and
urban employment have a point-estimate which is close in magnitude to the actual ones. It
is reassuring that a unique theoretical framework with 8 parameters is able to capture all the
qualitative interpretations one would get by running 5 different regressions, and also most
of the magnitudes.
The model does not necessarily have to fit the data, however, part of the fit can be
attributed to the minimization algorithm. Here, I provide two additional elements of the
predictive power of the model that are not directly related to the minimization procedure.
First, figure 12 shows the results from estimating the non-parametric relationship between
the dependent variables and distance to roads. The model does a reasonable job fitting the
non-linear relationship between the endogenous variables and distance to roads. To some
extent, this is an out-of-the-sample fit of the model. Even though the minimization algorithm
uses the coefficient of the regressions on distance to federal roads, this is not sufficient to
produce the non-linearity we observe in the figure. In a simple regression analysis, for
example, we would need to transform the explanatory variable (by taking the square or
the log) to generate a non-linear relationship. Second, note that not only the non-linear
relationship of the simulated dataset resembles the actual one, but also the noise in the
relationship between production variables TS and TC also follows each other. In the model,
this is generated only by the variation in the size of counties.
The best fit of the model was achieved at the parameters given in table 7. The parameter
φ affects the productivity of both cattle and soybeans. Different from previous estimation
in the Trade literature, it does not have any specific interpretation here. pS is estimated
to be 9 times higher than pL. This shows that soybeans are in fact more profitable than
cattle. However, the large fixed costs of soybeans with respect to cattle prevents farmer
22Part of the data also seemed to have absurd high levels of soybean production and urban activities.Drop these counties from the sample (around 7 counties are dropped out of 424).
28
from specializing in soybeans. The transportation costs for soybeans is also larger, which
is an additional force that pushes soybean production closer to roads. γM is estimated to
be very low. This is in accordance to figure 12, where we do not observe a large variation
in wages across different counties. In other words, there is little need for compensation on
wages. v is the parameter related to V (such that V = exp(v). The parameter αM indicates
that roughly 50% of income goes to imported goods. Finally, φ = 0.55 provides the labor
intensity factor. Mundlak et al. (1999) provides a review of the literature on the estimates
of intensity factors in agricultural production functions. Previous estimates of the labor
intensity varies widely, from 0.11 to 0.66, with the median estimate across articles being at
0.42. Therefore, it seems that the estimate of this parameter depends on the context, but it
is reassuring that the level found is in-between the range of parameters previously found in
the literature.
5.4 Counterfactual Analysis
Given the parameters estimated in the previous section, I follow with four counterfactuals
to learn about the mechanisms of the model. First, I increase the fixed costs of soybeans
production to make it inviable. Second, I proceed with a similar analysis for cattle. Third,
I increase the proximity of every county to federal roads in 10%, and, fourth, I decrease
the distance of counties to federal highways. The results for the counterfactual analysis are
presented in table 8.
When we exclude the possibility of soybean production, there is an expansion of cattle.
This is to be expected. Land that was being used for soybeans will now be used for cattle
production. Note, however, that the effect is not a large one, mainly because soybeans
were occupying a small share of the total land in each county. In spite of that, both urban
and rural employment almost disappear in the absence of the soybean sector. This result
shows the importance of the possibility of producing a highly profitable activity for the
urbanization process in the Amazon, and it also indicates why the Amazon did not engage
in the production of these crops before. In the absence of the commodity price shock, farmers
would not be able to pay for the large fixed costs of producing these activities.
Finally, the effect of varying distance to roads have a moderate effect on the production
of the region. This indicates that marginal changes in the infrastructure system may not
bring drastic changes in the production of the region.
29
6 Conclusion
Once the Brazilian federal government constructed a large highway system for the Amazon
in the early 1970s, millions of workers migrated to the region with the hope of producing
agricultural varieties that were common in their native cities. However, for almost 30 years
following the initial efforts of the government to populate the Amazon, the economy of the
region would be characterized by frustrated attempts to establish an agricultural industry.
In the middle of the 1990s, technological advances in seed made the production of grains
possible in the low latitude region of the Amazon, and the drastic increase in commodity
prices made the production of these commodities a highly profitable activity. Once these
conditions for the establishment of an export oriented agricultural industry came together,
farmers who were close to the federal highway system were disproportionately benefited by
the facility to export their products. As a consequence, a large urban sector rose to support
the increasing demand for goods generated by this export oriented sector.
The results from this paper provide three main contributions to the literature on develop-
ment and trade. First, this is one of the first studies to provide microeconomic evidence on
the mechanisms that may foster urbanization. According to the results from the article, the
absence of access to transportation, good prices for exporting goods, and the possibility of
producing mechanized crops may prevent the establishment of urban centers in rural settings
of developing countries. Second, recent articles highlight the role of mechanization and scale
for the low agricultural productivity in developing countries (Foster and Rosenzweig, 2011;
Restuccia and Adamopoulos, 2014; Adamopoulos and Restuccia, 2014). The present results
suggest that access to transportation infrastructure and the profitability or production may
constitute essential elements for the establishment of a large scale rural production. Third,
this article has shed light on the additional elements that may foster the benefits from invest-
ments in transportation infrastructure. In the Brazilian Amazon, it took more than 30 years
for the federal highway network to substantially affect its economic activities. The evidence
in the article suggests that the lack of the right macro-economic conditions for trade have
undermined the impact of roads in the region.
It is important to recognized that the Brazilian economy provides favorable conditions
for agricultural production that limits the external validity of this study. First, some re-
gions may not have access to seeds suited for the production of large-scale mechanized crops.
In Brazil, the availability of seeds suitable for low latitude areas took several years of re-
search financed by the federal government. Second, Brazil has a huge population that has
systematically migrated to take advantage of economic opportunities in several historical
30
moments of economic growth. This population represented an important source of labor for
the expansion of economic activities in the Amazon.
In spite of these advantages given by the Brazilian institutional setting, it is still interest-
ing how farmers in the Amazon region managed to expand their production of export-oriented
commodities by more than 300% in roughly 15 years. The production of grains is a large-
scale enterprise and its production requires technology, large investments in machinery, and
access to international markets, which makes this region a specially good case for studying
the emergence of large-scale agricultural enterprises in developing countries. Here, I focused
on the role of transportation infrastructure for the development of agriculture. In future
research, I intend to explore additional conditions, such as access to credit and technology,
and the role they might play in the expansion of large scale agricultural production.
31
References
Tasso Adamopoulos and Diego Restuccia. The size distribution of farms and international
productivity differences. American Economic Review, 104(6):1667–97, 2014.
Elizabeth Barona, Navin Ramankutty, Glenn Hyman, and Oliver T Coomes. The role of
pasture and soybean in deforestation of the brazilian amazon. Environmental Research
Letters, 5(2):024002, 2010.
Nathaniel Baum-Snow. Did highways cause suburbanization? The Quarterly Journal of
Economics, 122(2):775–805, 2007.
Paula Bustos, Bruno Caprettini, and Jacopo Ponticelli. Agricultural productivity and struc-
tural transformation: Evidence from brazil. Working Paper, 2013.
Jonathan Eaton and Samuel Kortum. Technology, geography, and trade. Econometrica, 70
(5):1741–1779, 2002.
Benjamin Faber. Trade integration, market size, and industrialization: Evidence from china’s
national trunk highway system. The Review of Economics Studies, 2014.
Pablo Fajgelbaum and Stephen Redding. External integration, structural transformation,
and economic development: Evdence from argentina 1870-1914. NBER Working Paper,
20217, 2014.
Andrew D Foster and Mark R Rosenzweig. Agricultural productivity growth, rural economic
diversity, and economic reforms: India, 1970–2000*. Economic Development and Cultural
Change, 52(3):509–542, 2004.
Andrew D Foster and Mark R Rosenzweig. Economic development and the decline of agri-
cultural employment. Handbook of development economics, 4:3051–3083, 2007.
Andrew D Foster and Mark R Rosenzweig. Are indian farms too small? mechanization,
agency costs, and farm efficiency. Econojmic Growth Center, Yale University New Haven
CT, 2011.
Douglas Gollin, Remi Jedwab, and Dietrich Vollrath. Urbanization without structural trans-
formation: Evidence from sub-saharan africa. 2013.
32
J Vernon Henderson, Mark Roberts, and Adam Storeygard. Is urbanization in sub-saharan
africa different? 2013.
Berthold Herrendorf, Richard Rogerson, and Akos Valentinyi. Growth and structural trans-
formation. 2013.
Remi Jedwab. Urbanization without structural transformation: Evidence from consumption
cities in africa. Processed, George Washington University, 2013.
Simon Kuznets. Modern economic growth: findings and reflections. The American Economic
Review, pages 247–258, 1973.
Robert E Lucas. On the size distribution of business firms. The Bell Journal of Economics,
pages 508–523, 1978.
Douglas C Morton, Ruth S DeFries, Yosio E Shimabukuro, Liana O Anderson, Egidio Arai,
Fernando del Bon Espirito-Santo, Ramon Freitas, and Jeff Morisette. Cropland expansion
changes deforestation dynamics in the southern brazilian amazon. Proceedings of the
National Academy of Sciences, 103(39):14637–14641, 2006.
Yair Mundlak, Donald F Larson, and Rita Butzer. Rethinking within and between regres-
sions: The case of agricultural production functions. Annales d’Economie et de Statistique,
pages 475–501, 1999.
Alexander Pfaff, Juan Robalino, Robert Walker, Steven Aldrich, Marcellus Caldas, Eu-
staquio Reis, Stephen Perz, Claudio Bohrer, Eugenio Arima, William Laurance, et al.
Road investments, spatial spillovers, and deforestation in the brazilian amazon. Journal
of Regional Science, 47(1):109–123, 2007.
Stephen J Redding and Matthew A Turner. Transportation costs and the spatial organiza-
tion of economic activity. Chapter in progress for the Handbook of Regional and Urban
Economics, 2014.
Diego Restuccia and Tasso Adamopoulos. Land reform and productivity: A quantitative
analysis with micro data. Working Paper, 2014.
Peter D Richards, Robert J Myers, Scott M Swinton, and Robert T Walker. Exchange
rates, soybean supply response, and deforestation in south america. Global Environmental
Change, 22(2):454–462, 2012.
33
Jennifer Roback. Wages, rents, and the quality of life. Journal of Political Economy, 1982.
Peter M Robinson. Root-n-consistent semiparametric regression. Econometrica: Journal of
the Econometric Society, pages 931–954, 1988.
Adam Storeygard. Farther on down the road: transport costs, trade and urban growth in
sub-saharan africa. 2013.
Maria del Carmen Vera-Diaz, Robert K Kaufmann, Daniel C Nepstad, and Peter Schlesinger.
An interdisciplinary model of soybean yield in the amazon basin: the climatic, edaphic,
and economic determinants. Ecological Economics, 65(2):420–431, 2008.
Robert Walker, John Browder, Eugenio Arima, Cynthia Simmons, Ritaumaria Pereira, Mar-
cellus Caldas, Ricardo Shirota, and Sergio de Zen. Ranching and the new global range:
Amazonia in the 21st century. Geoforum, 40(5):732–745, 2009.
34
Figure 1: Historical Population of the Brazilian Amazon.
Rubber Cycle
National Plan ofIntegration
Commodity PriceIncrease
.04
.06
.08
.1.1
2.1
4A
maz
on's
sha
re o
f Bra
zilia
n po
p.
05
1015
20T
otal
pop
. (in
mill
ions
)
1872
1890
1900
1910
1920
1940
1950
1960
1970
1980
1991
2000
2010
Year
Total Amazon population Amazon's share of Brazilian pop
35
Figure 2: Soybean price at local retailers and the international soybean price adjusted byexchange rate.
1020
3040
Loca
l Pric
e
200
400
600
800
1000
Adj
uste
d in
tern
atio
nal p
rice
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Year
Adjusted international price Local Price
Figure 3: Trends in the agricultural production in the Brazilian Amazon.
2000
030
000
4000
050
000
6000
0P
rodu
ctio
n (in
thou
sand
s)
2000
4000
6000
8000
1000
0In
ha
(in th
ousa
nds)
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Soy Other crops Cattle
36
Figure 4: Trends in employment by sector in the Brazilian Amazon (in thousands).
010
020
030
040
050
0E
mpl
oym
ent
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Urban Rural
Figure 5: Trends in main variables over time.
-2-1
01
2N
orm
aliz
ed v
aria
ble
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Cattle Soy Urban JobsRural Jobs Adjusted price
Notes: The variables used to produce this figure were normalized according tothe mean the the standard deviation across the years.
37
Figure 6: County Boundaries, Project Lines, and Constructed Roads.
LegendCapitalsConstructed RoadsProject Lines
0 370 740 1,110 1,480185Kilometers
¯
38
Figure 7: Difference between counties close and far from federal roads in terms of populationand income.
050
0010
000
1500
0D
iffer
ence
in In
com
e
010
000
2000
030
000
4000
0D
iffer
ence
in P
opul
atio
n
1950
1960
1970
1980
1990
2000
2010
Population Income
Notes: This figure reports the difference in the average population andtotal income between counties closer than 100 km from a constructedfederal road and those farther away.
Figure 8: Difference between counties close and far from federal roads in terms of commodityproduction.
-400
00-2
0000
020
000
4000
060
000
Diff
eren
ce in
Cat
tle
5000
1000
015
000
2000
025
000
3000
0D
iffer
ence
in S
oy
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Soy Cattle
39
Figure 9: Heterogeneity in cattle production: difference between counties close and far fromfederal roads in terms of their production of soybeans in 2010.
5010
015
020
0C
attle
pro
duct
ion
(in th
ousa
nds)
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Large Medium Small
Notes: Large producers are the counties that reached 2010 abovethe 95th percentile in terms of soybean production. Medium arethose between the 95th and the 90th percentile. And small pro-ducers are the rest of the counties.
Figure 10: Difference between counties close and far from federal roads in terms of employ-ment in the urban and rural sectors.
5010
015
020
025
030
0D
iffer
ence
in R
ural
050
010
0015
0020
00D
iffer
ence
in U
rban
1990
1995
2000
2005
2010
Urban Rural
A. Different Axis
050
010
0015
0020
00D
iffer
ence
1990
1995
2000
2005
2010
Urban Rural
B. Same Axis
40
Figure 11: Non-parametric regressions of the differential trend in the main endogenousvariables of the model.
-200
00-1
0000
010
000
Diff
eren
ce
0 50 100 150 200 250 300Distance to roads
Before price shock (1999-1994)
After price shock (2004-1999)
A. Soybeans and related crops
-400
00-2
0000
020
000
4000
0D
iffer
ence
0 50 100 150 200 250 300Distance to roads
Before price shock (1999-1994)
After price shock (2004-1999)
B. Cattle
-500
050
010
0015
0020
00D
iffer
ence
0 50 100 150 200 250 300Distance to roads
Before price shock (1999-1994)
After price shock (2004-1999)
A. Urban Sector
-500
050
010
0015
0020
00D
iffer
ence
0 50 100 150 200 250 300Distance to roads
Before price shock (1999-1994)
After price shock (2004-1999)
B. Rural Sector
Notes: The non-parametric regressions were adjusted by a cubic polynomial on distance to roadsfollowing the procedure suggested in Robinson (1988).
41
Figure 12: Non-parametric regressions of the endogenous variables in the model on distanceto roads (actual and simulated dataset).
A. Soy
0 50 100 150 200 250 3000
10
20
30
40
50
60
70
80
90
100
DataSimulation
B. Cattle
0 50 100 150 200 250 300200
400
600
800
1000
1200
1400
1600
1800
2000
2200
DataSimulation
C. Rural
0 50 100 150 200 250 3000
200
400
600
800
1000
1200
1400
1600
1800
2000
DataSimulation
D. Urban Sector
0 50 100 150 200 250 3000
200
400
600
800
1000
1200
1400
1600
1800
2000
DataSimulation
E. Wages
0 50 100 150 200 250 300500
600
700
800
900
1000
1100
1200
1300
1400
1500
DataSimulation
42
Table 1: Descriptive Statistics
1950-2010 dataset 1990-2010 datasetElasticity wrt Elasticity wrt
Roads RoadsFull No State Full No State
Mean Sample Capitals Mean Sample Capitals(1) (2) (3) (4) (5) (6)
Panel A. Geographic CharacteristicsArea 35206.90 0.80* 0.80* 12825.72 0.46* 0.46*
(51082.49) (0.03) (0.02) (19321.80) (0.09) (0.09)Corn suitability 5.30 0.01 0.01 5.20 -0.01 -0.01
(0.53) (0.01) (0.01) (0.56) (0.01) (0.01)Soybeans suitability 4.20 -0.03 -0.03 4.37 -0.04 -0.04
(0.84) (0.02) (0.02) (0.98) (0.02) (0.02)Slope 4.37 0.23 0.25 2.70 0.07 0.07
(7.70) (0.31) (0.34) (5.91) (0.19) (0.20)Distance to rivers 34.53 0.07 0.06 34.95 0.06 0.06
(22.30) (0.08) (0.08) (25.64) (0.06) (0.06)Temperature in september 267.60 0.01* 0.01* 263.07 0.01 0.01
(8.78) (0.01) (0.01) (11.63) (0.01) (0.01)Rainfall in september 71.83 0.06 0.07 72.62 0.06 0.06
(26.04) (0.05) (0.06) (26.72) (0.03) (0.03)
Panel B. Initial ConditionsPopulation in 1950 18504.99 -0.07 -0.02 - - -
(28226.78) (0.05) (0.05)Population in 1960 26126.17 -0.10* -0.05 - - -
(43720.05) (0.04) (0.04)
Panel C. Determinants of RoadDistance to capitals 248.43 0.53* 0.53* 300.98 0.27* 0.26*
(169.93) (0.05) (0.05) (202.29) (0.03) (0.03)Distance to project lines 86.18 0.64* 0.63* 82.09 0.50* 0.50*
(56.44) (0.11) (0.12) (6.01) (0.07) (0.07)
Obs
Notes: * p<0.5. Robust standard errors at the State level in parentheses. Columns 2 to 5 presents coefficientsof regressions of the log of each variable against the log of distance to roads. Column 3 drops 7 State capitalswhich where nodal points in the Federal government project.
43
Table 2: Effect of access to federal roads on population and income by period.
Period2004-1994 2004-1999 (2004-1999)-(1999-1994)
(1) (2) (3) (4) (5) (6)Panel A. PopulationClose 14,922 11,785 26,593*** 30,420** 11,671 18,635*
(9,402) (9,748) (9,942) (13,301) (7,192) (10,765)Baseline 22,134 - 52,536 - - -P-value - 0.001 - 0.068 - 0.26
Panel C. IncomeClose 1,509 461.1 18,446** 15,560* 15,046*** 14,602**
(1,287) (1,141) (7,650) (8,126) (5,369) (6,261)Baseline 889 - 4,773 - - -P-value - 0.001 - 0.001 - 0.001
Dist to Capitals Y Y Y Y Y YRegion FE N Y N Y N YGeography N Y N Y N YInitial Condition N Y N Y N Y
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors at the county level in parentheses.Every regression controls for county and year fixed effect. Dist to Capitals is a cubic polynomialon the distance of each county to the State Capital. Geography represent controls for: area ofthe county, corn suitability, soybeans suitability, slope, temperature in September, and rainfall inseptember. Initial condition controls for the population in 1950 and 1960. Region FE stands fora dummy wether the county is the southern part of the Amazon (the States of Mato Grosso orRondonia). Sample includes 99 counties. Baseline stands for the mean for the counties far fromroads in the initial period. P-value is the result from a joint test of the statistical significance ofthe control variables.
44
Table 3: Effect of access to federal roads on agricultural production by period.
Period2004-1994 2004-1999 (2004-1999)-(1999-1994)
(1) (2) (3) (4) (5) (6)Panel A. Soybeans and related crops (in ha)Close -1,517 -2,662** 10,979* 11,541* 12,496*** 14,202**
(2,070) (971.2) (5,161) (4,923) (3,284) (5,251)Baseline 3,620 - 4,366 - - -P-value - 0.001 - 0.001 - 0.001
Panel B. Other crops (in ha)Close -3,845 -3,523 1,218 2,425 1,218 2,425
(3,108) (3,053) (2,401) (2,787) (2,401) (2,787)Baseline 5,199 - 6,801 - - -P-value - 0.001 - 0.001 - 0.001
Panel C. Cattle (output)Close -17,383 -14,765 -6,276 -21,195 11,108 -6,430
(11,207) (7,740) (36,091) (21,256) (30,308) (25,964)Baseline 111303 - 138071 - - -P-value - 0.001 - 0.053 - 0.053
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors at the state level inparentheses. Controls are explained in table 2. Sample includes 259 counties.
Table 4: Effect of access to federal roads on employment in the urban and rural sector byperiod.
Period2004-1994 2004-1999 (2004-1999)-(1999-1994)
(1) (2) (3) (4) (5) (6)Panel A. Urban SectorClose 111.0 154.0 648.6*** 507.3*** 537.6*** 353.3*
(134.7) (152.1) (153.7) (161.3) (206.4) (213.0)Baseline 272 - 321 - - -P-value - 0.082 - 0.080 - 0.167
Panel B. Rural SectorClose 54.56** 21.48 147.7*** 93.87* 93.16* 72.39
(26.16) (28.40) (43.08) (53.04) (52.72) (67.09)Baseline 51 - 75 - - -P-value - 0.001 - 0.001 - 0.016
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors at the state level inparentheses. Controls are explained in table 2. Sample includes 259 counties.
45
Table 5: Robustness check.
Dependent Variables and Period(2010-1991)-(1991-1970) (2004-1999)-(1999-1994)
Pop Inc Soy Urban Rural(1) (2) (3) (4) (5)
Panel A. Baseline ResultsClose 11,671 15,046** 12,496*** 573.6*** 93.16*
(7,192) (5,369) (3,284) (206.4) (52.72)Obs 99 99 259 259 259
Panel B. Dropping Service CountiesClose 15,374** 13,243** 11,917* 389.3*** 33.81
(7,274) (5,796) (5,658) (93.63) (85.93)Obs 80 80 205 205 205
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors at the statelevel in columns 3-5 and at the county level in columns 1 and 2 in parentheses.All regressions control for a cubic polynomial of the distance a county is froma State Capital.
46
Table 6: Fit of the model
Panel A. Endogenous Variables (Mavg)TS TC LA LU w
Data 26 914 495 673 861Model 29 912 490 675 860
Panel B. Regressions on Roads (Mβ)βTS
βTcβLa
βLuβw
Data -0.12 2.10 -0.08 -2.66 -0.24Model -0.10 7.61 -1.62 -2.28 -0.04
Table 7: Estimated parameters
θ pS/pC fS/fC γS/γC γM v αM φ31.57 9.62 679.2 1.289 -0.0001 3.28 0.48 0.55(2.04) (1.12) (208) (0.12) (0.00001) (0.11) (0.01) (0.01)
Notes: Bootstrapped standard errors in parentheses (50 samples). PCand fC are equal to 10.
Table 8: Counterfactuals
TS TC LA LU wModel 29 912 490 675 860No soy 0 942 31 48 860Closer to Roads (10%) 32 929 532 732 860Farther from Roads (10%) 27 895 454 625 860
47
A The Plans of Integration of the Brazilian Amazon
Figure A1: The National Plan of Integration (1969).
Figure A2: The National Transportation Plan (1973).
48
B Datasets
B.1 NASA’s Shuttle Radar Topography Mission
I use topographic data produced as part of NASA’s Shutte Radar Topographic Missions.
I use the 3arc second scale (around 90 m2) to calculate the standard deviation in altitude
in each city as our measure of slope. For the altitude we used the average altitude in each
municipality.
B.2 FAO-GAEZ
To obtain information on soy and corn suitability, I use data from the Global Agro-Ecological
Zones database produced by the FAO. The suitability is measured as the potential yields
attainable for a crop in a certain geographical area. They use information on climatic
conditions, slope, and the level of technology available to produce a raster file of potential
yields for different levels of technology. I included the high input type of technology, which
corresponds to the large-scale mechanized techniques that are used in the region. I then
measure the averate suitability in each municipality.
B.3 RAIS
For employment I used data from the Relacao Anual de Informacoes Sociais (RAIS). This
dataset is provided by the Ministry of Labor from their official recors on all the formal job
employments in Brazil. To obtain the dataset, I had to require access to a website that
compiles the information at the county level in each year since the 1980s. However, the
way the website is organized required a webscrapping technique to automatize the download
of the information. I used the internet language iMacro that has been recently developed
specially for this type of procedure.
B.4 World Climate
For information in temperature and precipitation we used the WorldClim dataset available
at http://www.worldclim.org/. The WorldClim data is a set of global climate layers (climate
grids) with a spatial resolution of 1 square kilometer. The dataset provides montly precipita-
tion and temperature averages using information from 1950 to 2000. For each municipality,
calculated the average monthly precipitation and temperature in each month of the year.
49
B.5 IMEA
Data for local retailer’s price of soybeans and cattle was obtained from the Instituto Mato
Grossense de Economia Agropecuaria. This a public research institute that is part of the
State government of Mato Grosso. They provide daily price offered by retailer’s in several
cities for soybeans from 1996 to 2013.
Table B1: Summary of Datasets
Dataset Institution Information Frequency/Years(1) (2) (3) (4)
Dataset 1: 1950-2010CENSUS IBGE (IPEA) Population DecadalCENSUS IBGE (IPEA) Income Decadal (since 1970)
Dataset 2: 1991-2010RAIS Ministry of Labor Labor per sector AnnualPAM IBGE (IPEA) Land use per activity AnnualPPM IBGE (IPEA) Cattle production Annual
Both datasets: geographic characteristics, shapefiles and othersGAEZ FAO Soy and corn suitabilitySRTM NASA Slope and elevation
WorldClim WorldClim Rainfall and temperatureCities map CENSUS Area and political boundariesRoad map IBAMA Road networkRivers map IBAMA Principal rivers
Agro CENSUS IBGE Number of tractors 2007
Notes: (IPEA) indicates that, even though the primary source of information is collected by another institu-tion, the information for this article was organized and published in the websited of the Insituto de Pesquisae Economic Aplicada (IPEA), a public research institute of the federal government.
50
C Comparative Statics
Figure C1 summarizes the main relationships between the endogenous variables in the model
(population, income, land use and labor employment) to the exogenous variables. Panel A
shows the production characteristics before the shock on commodity prices. Panel B shows
the production after a positive shock. As in section 3 of the paper, I divide the discussion
in two.
Variation in Distance to Roads given International Price
Panel A figure B shows a monotonic relationship between the land area used for soybean
and distance to roads. However, cattle follows a U-shape relationship with distance to roads.
In the first few kilometers, there is a positive relationship, but after 100 km the relationship
is clearly negative. As discussed in the main body of the article, this can be attributed to
two margins of substitutions: the one between cattle and soybeans, and the one between
cattle and forests. Given the parameters of the model, employment is very low and there is
an almost inexistent urban center.
Variation in International Prices given Distance to Roads
Panel B shows the results from simulations when I increase commodity prices by 100%.
First, the simulation predicts one main stylized fact from the literature: that soybean pro-
duction precedes cattle ranching. Counties close to federal roads who were initially larger
producer of cattle expand their production of soybeans over pasture. They become smaller
producers when compared to counties at a medium distance to roads (100-200 km). The
expansion of cattle in counties farther from roads leads to a large deforestation. Note, for
example, that a county at 150 km from a road has an expansion of roughly 200 land units in
the production of agricultural goods (from 200 to 400), while counties at 0 km from a road
have an expansion of around 150. Depending on the parameters of the model, this could be
magnified.
51
Figure C1: Relationship between the endogenous variables in the model and distance toroads.
0 100 200 3000
10
20
30
40
50
60
70
Distance to roads
A. Labor Employment
L
R
LU
Panel A. Low Price
0 100 200 3000
100
200
300
400
500
600
700
Distance to roads
B. Land Employment
T
C
TS
0 100 200 3000
10
20
30
40
50
60
70
Distance to roads
A. Labor Employment
L
R
LU
Panel B. High Price
0 100 200 3000
100
200
300
400
500
600
700
Distance to roads
B. Land Employment
T
C
TS
Notes: Parameters used for the simulation are: pM = 10, φ = 0.1, fS = 15,fC , αU = 0.5, V = 5, aS = aC = aU = 1, θ = 4, µ = 1, γ = 0.005. µ is thelevel parameter of the Frechet. For low prices, in panel A, we have pS = 7.5and pC = 1.5. In panel B prices are doubles.
52
D Additional Evidence of the Impact of Commodity
Prices on Production
As discussed in section 2 of this article. I argue that the devaluation of the exchange rate
between 1999 and 2004 had a large impact on the price of export oriented product faced
by farmers in the Amazon. As a consequence, this had a large impact on the profitability
of the rural activities which led to a large expansion in production. One natural question
is whether the increase in agricultural production was specific to the products exported in
the amazon region. If that were the case, this would suggest that, besides the price shock,
there were additional technological improvements in production that may raise questions
on the interpretation of the results. In particular, it would cast doubts on the established
association between commodity production and commodity prices.
Figure D1 presents the results from an analysis of several commodities that are recognized
as important for the Brazilian export basket. Each variable in figure D1 is the total amount
produced in Brazil normalized by its mean and standard deviation across the years. To
facilitate the visualization, I divide the set of commodities in two with no specific rule.
Panel A in figure D1 presents the first set and shows that both soybeans and cattle follows
the international commodity price increase adjusted by the exchange rate. However, tobacco,
which is primarily produced in the south of Brazil, has a dramatic increase during this period.
Panel B registers the expansion of two additional products during this period: wheat and
grapes. Both of them are primarily produced in the south of Brazil. It is interesting to note
that sugarcane, which is largely produced in the southeast, did not follow the variation in the
exchange rate during the period. One of the main final goods from sugarcane between 1999
and 2003 was the production of ethanol. During this period, there was a massive adoption of
bi-fuel technologies in Brazilian cars and, as a strategy to reduce the effect of the exchange
rate devaluation on prices, the government controlled the price of fuel and ethanol during
the period (this has been a common strategy from the government to influence inflation in
the past two decades). When the international price of oil increased after 2005, however, the
government allowed the price of ethanol and fuel increase, which could have motivated the
expansion in sugarcane production. To some extent, the different behavior in the production
of sugarcane is an exception that confirms the rule.
Finally, I obtained aggregate information on certain exporting agricultural products in
Brazil from 1995 to 2004. Figure D2 presents the trend in the normalized exporting in a
set of commodities with available information. Data is given in FOB dollars. Note that,
53
between 1995 and 2004, the price of commodities were relatively stable in the international
markets. Therefore, the expansion in trends can be attributed to an increase in the total
amount of exported commodities. Trends in all the commodities in the figure are very similar
and follow the international price of soybeans adjusted by the exchange rate. This supports
the price of soybeans as a good proxy for the price of commodity in general.
Figure D1: Trends in the production of a set of relevant export oriented commodities inBrazil.
-2-1
01
2N
orm
aliz
ed v
aria
ble
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Cattle Soy Tobacco Price
A. First Set
-2-1
01
2N
orm
aliz
ed v
aria
ble
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Grapes Wheat Sugarcane Price
B. Second Set
54
Figure D2: Trends in the exporting of a set of relevant commodities in Brazil (FOB data).
-10
12
3N
orm
aliz
ed v
aria
ble
1994
1996
1998
2000
2002
2004
Cattle Soy TobaccoGrapes Total Adj Commodity Price
55
E Technological Characteristics of the Agricultural Pro-
duction
In section 2 of this article I discuss the complementarity and substitutability between the
main agricultural varieties produced in the Amazon region. Here, I provide data evidence
to complement the discussion.
Figure E1 presents evidence of complementarity between soybeans, corn, and cotton, but
a substitutability between soybeans and cattle. Further discussion is included in the main
article.
Figure E1: Complementarity and Substitutability between the Main Agricultural Activities.
0.2
.4.6
.8C
attle
Are
a
0 .2 .4 .6 .8Soy Area
Data
.04
.06
.08
.1.1
2C
attle
Are
a
0 .2 .4 .6 .8Soy Area
Non-parametric Regression
0.1
.2.3
.4.5
Cor
n A
rea
0 .2 .4 .6 .8Soy Area
Data
0.1
.2.3
Cor
n A
rea
0 .2 .4 .6 .8Soy Area
Non-parametric Regression
0.0
5.1
.15
Cot
ton
Are
a
0 .2 .4 .6 .8Soy Area
Data
0.0
1.0
2.0
3.0
4C
otto
n A
rea
0 .2 .4 .6 .8Soy Area
Non-parametric Regression
Table E1 presents evidence of larger fixed costs in the production of cattle with respect
to soybeans.
56
Table E1: Mechanization of Soybeans and Cattle
Dependent variableSmall Tractors Big Tractors
(1) (2)Log of cattle 9.473*** 4.332**
(1.613) (1.747)Log of soybeans 5.601*** 14.58***
(1.196) (2.065)
Obs 225 225R2 0.348 0.491
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard inparentheses. Data from Agricultural Census of 2007. Big tractorshave more than 100 horsepower.
57
F Model Details
F.1 Equations defining the equilibrium
Consumers
Assume consumers have a cobb-douglas preference:
U = δcαMM cαUU (14)
Where δ =1
ααMM ααUU. cM are the goods imported from other regions:
Demand is given by:
cMi = αMwipM
cUi = αUwipU
(15)
Where i indexes the city. Workers satisfy an indirect utility that provides the non-
migration condition.
V =wi
eγαMDroadiPi(16)
Where the price index is equal to P = (pαMM pαUUi ). We will see that, given the firm profit
maximizing conditions, pUi = wi/aU . Droads is the distance from roads. The indirect utility
function can be rewritten as:
V =wi
eα1γDroadipαMM pαUU
=w1−αUi aαUU
eγαMDroadipαMM
(17)
Firms
For production, there will be 2 technologies for agricultural production:
y(`)S = z(`)l(`)φS (18)
y(`)C = z(`)l(`)φC (19)
58
Where ` defines the plot which are distributed according to a measure of land Ti. There-
fore, ` ∈ (0, Ti]. Also, note that the production will depend on the relative price of soybeans
and cattle. Here, I abstract from potential differences in total factor productivity as I am
not able to disentangle these from the price (in other words, pSaS and pCaC are combined
into one variable). Each plot has a realization of an extreme value distribution (gumbel type
II) 23.
The maximization problem, if producing the modern crop, is:
max pSe−γSDroadiz(`)l(`)φ − wil(`)− fS (20)
For cattle, the maximization problem is:
max pCe−γCDroadiz(`)l(`)φ − wil(`)− fC (21)
And we have the following maximization conditions.
wi = pUaU
wi = pCe−γSDroadiz(`)φl(`)φ−1
wi = pSe−γCDroadiz(`)φl(`)φ−1
wi > 0
(22)
The last condition is given by an inequality for the plots with no production.
Potential profits in each plot is given by:
π(`)S = (pSe
−γDroadiz(`)
wφi)
11−φΦ− fS
π(`)C = (pCe
−γDroadiz(`)
wφi)
11−φΦ− fC
π(`) = 0
(23)
Where Φ = (φφ
1−φ − φ1
1−φ ) The landowner can choose to not use the land. If he decides
23I chose this distribution because it is readily available in Matlab.
59
not to clear the land, profits are equal to zero.
The production is given by the following equations:
YSi =
(pSφ
wi
) φ1−θ ∫∞
ziz
11−φdz
YCi =
(pCφ
wi
) φ1−θ ∫ zi
ziz
11−φdz
Labor Market Clearing
Note that, because this is an open city model, Ni is endogenous.
LSi + LCi + LUi = Ni (24)
Where l(`)C and l(`)S are given by the following in each plot of land with positive
production (labor demand is zero for uncleared land):
l(`)S = (pSe
−γSDroadsiφz(`)
wi)
11−φ
l(`)C = (pCe
−γCDroadsiφz(`)
wi)
11−φ
(25)
Total labor in the soybean sector is given by the following (where T Si is the total area
producing soybeans and TCi the one producing cattle) equation. Also, assume that each city
has a measure Ti of land:
∫`∈TS
l(`)Sd` = (pSe
−γSDroadsiφ
wi)
11−φ∫`∈TSi
z(`)1
1−φd`
LSi = (pSe
−γCDroadsiφ
wi)
11−φ∫`∈TSi
z(`)1
1−φd`
(26)
And cattle is obtained analogously:
LCi = (pCe
−γCDroadsiφ
wi)
11−φ∫`∈TCi
z(`)1
1−φd` (27)
In the service sector, labor will be defined by the non-tradeable sector, which means that
the demand for labor from the landowner and the workers will have to meet the supply of
60
goods.
Non-tradeable condition
αUIi = pUiyUi
αUIi = wiLUi
(28)
And the total income is equal to:
Ii = Profits+Wages
Ii =∫`∈TS
πS(z(`))d`+∫`∈TC
πC(z(`))d`+ wiNi
(29)
Tradeable condition
For the tradable balance we need the following:
αMIi = Soybeansi + Cattlei
αMIi =∫`∈TSi
pSe−γSDroadsiyS(z(`))d`−
∫`∈TSi
fSd`
+∫`∈TCi
pCe−γCDroadsiyC(z(`))d`−
∫`∈TCi
fCf(z)dz(`)
αMIi =∫`∈TSi
wil(`)Sφ
d`− fST Si
+∫`∈TCi
wil(`)Cφ
d`− fCTCi
αMIi =wiφ
(LSi + LCi)− T Si fS − TCi fC
(30)
Note that we use the following transformation here.
61
wi = pxe−γxDroadsiz(`)l(`)φ−1φ
li(`)wiφ
= pxe−γxDroadsiz(`)l(`)φ
li(`)wiφ
= pxe−γxDroadsiyx(`)
(31)
Where x ∈ {S,C}. Combining both the tradeable and the non-tradeable sector (and
aggregating the labor demand in the agricultural sector), we have:
LSi =αUαM
(LAiφ− T Si
fSwi− TCi
fCwi
)(32)
LAi is the total employment in agriculture. Note that, in the absence of roads, we would
have the same proportion of people employed in the service sector and people employed in
the agricultural one in every city. However, because of the fixed costs, we do not have the
exact relationship.
Equilibrium
With the following set of exogenous parameters
Σ = (γS, γC , θ, pS, pC , pU , aU , φ, V, fS, fC , αU , αS)
, the equilibrium is then given by:
• Landowners maximize profits
• Firms maximize profits
• Agents maximize utility
• Agents are indifferent between different cities
• Labor market clears
• Balance in tradeable and non-tradeable sector
The equilibrium provides endogenously Xi = (LSi , LCi , L
Ui , T
Si , T
Ci , wi, p
Ui , Ni)
Equations defining the equilibrium
62
With the following set of exogenous parameters:
Σ = (γ, θ, pS, pC , pU , aU , φ, V, fS, fC , αM , αU)
, the equilibrium is then given by following equations:
LSi = (pSe
−γSDroadsiφ
wi)
11−φ
∫`∈TSi
z(`)1
1−φd` (33)
LSi = (pSe
−γSDroadsiφ
wi)
11−φ
∫`∈TSi
z(`)1
1−φd` (34)
LCi = (pCe
−γCDroadsiφ
wi)
11−φ
∫`∈TCi
z(`)1
1−φd` (35)
LUi =αUαM
(LAiφ− T Si
fSwi− TCi
fCwi
)(36)
wi = (V eγαMDroadipαMM
aαUS)
11−αU (37)
pUi =wiaU
(38)
YSi =
(pSe
γSDroadsiφ
wi
) φ1−φ
Ti
∫`∈TSi
z1
1−φd` (39)
YCi =
(pCe
γCDroadsiφ
wi
) φ1−φ
Ti
∫`∈TCi
z1
1−φd` (40)
The equilibrium provides endogenously Xi = (LSi , LCi , L
Ui , T
Si , T
Ci , wi, p
Ui , Ni, YSi, YCi)
F.2 Proof of Proposition (Partition of Land)
So far, I have worked with an arbitrary partition of land T Si and TCi . Here, I define the plots
of land in the county with cattle, soybeans, or forest according to quality of the plot z(`).
In the main body of the article, I stated the following proposition, which I prove below.
Proposition 2 (Partition of Land 1) Define zSCi as the plot (`) where the farmer would
be indifferent between producing cattle and soybeans, and zCFi for the plot (`) where the
63
farmer would be indifferent between producing cattle and clearing the forest. If the conditions
pS > pC, fS > fC, and zCFi < zSCi are satisfied, then the area of the county can be partitioned
in three where soybeans is produced in ` such that zCFi < z(`) < zCSi , cattle is produced in
zCFi < z(`) < zCSi , and forest is left uncleared for 0 < z(`) < zCFi .
Proof First, let’s define zCFi , zCSi , and zSFi . From the profit function, we can establish the
condition such that πS(`) ≥ πC(`) that will provide an equation for zCSi . Manipulation of this
inequality leads to:
z(`) ≥ (wφi
pSi − pCi)(fS − fC)1−φΦφ−1 (41)
Where Φ = (φφ
1−φ − φ1
1−φ ). An increase in z(`) always makes this inequality more likely
to be satisfied, which defines zCSi as an unique value of z(`) that satisfies the above equation
as an equality. We can derive zCFi from πC(`) ≥ 0, which leads to the following:
z(`) ≥ eγDroadsiwφip∗Ci
(fCΦ
)1−φ (42)
Again, there is only one zCFi that makes the above equation an equality 24. We can also
define zSFi in the same manner.
First, note, given pS > pC , better land will always be used for soybeans, because of the
complementarity between the quality of land zi and the price of the activity pS and pC . In
other words, the cross-derivative between zi and pS or pC is positive. As a consequence, we
know that for zi > zCSi , the land will be more profitable producing soybeans. Note also that,
because pS > pC and fS > fC , then zCSi > 0 in equation 41. Similar logic can be used to
show that zCFi > 0 and zSFi > 0.
Now, assume that zCFi < zSCi . If zSFi < zCFi , then it means that at the plot zSFi produc-
ing cattle leads to negative profits, while producing soybeans leads to a non-negative one.
However, if this is true, for the plot of land zSCi , the farmer would not be indifferent between
producing cattle and soybeans, because of zSFi < zSCi and the positive cross-derivative be-
tween zi and pS and pC , at plot zSCi the farmer would actually have a larger profit producing
soybeans, which contradicts the definition of zSCi . If zSFi > zCSi , it would mean that at the
plot zCSi the farmer has a negative profit (again, I use the positive cross-derivative), which
contradicts the fact that the farmer already has a positive profit with cattle at that point.
24Note that here the assumption that both cattle and soy have the same φ plays an important role infacilitating the solution of the problem. With different φ, we would have multiple solutions. However, thesame patterns we will look at the data would be possible to be generated by using different φ.
64
Finally, analysing zCFi < zSCi provides the parameter condition for the partition:
(fS − fCfC
)1−φ >pSi − pCipCi
(43)
This last inequality concludes the proof. �
Proposition 3 (Partition of Land 2) If the conditions pS > pC, fS > fC, and zCFi ≤zSCi are satisfied, then the area of the county can be partitioned in two where soybeans is
produced in ` such that zSFi < z(`), and forest is left uncleared for 0 < z(`) < zSFi . No cattle
is produced.
Proof Above zSCi , it is always more profitable to produce with soybeans. If zCFi ≤ zSCi , it
means that at the point when cattle becomes profitable for production, soybeans are already
a more profitable activity. Therefore, no cattle is produced. �
Proposition 4 (Partition of Land 3) If the conditions pS > pC and fS ≤ fC are sat-
isfied, then the area of the county can be partitioned in two where soybeans is produced in `
such that zSFi < z(`), and forest is left uncleared for 0 < z(`) < zSFi . No cattle is produced.
Proof If the price of soybeans is better than the one for cattle, and on the top of that its
fixed costs are smaller, then soybeans is always more profitable than cattle production. In
this case, the county specialized in soybean production. �
If we follow analogously with pC > pS, then we have the cases with cattle being produced
in the better plots within the county and another 3 partitions of land are obtained.
F.3 Revisiting equations defining the equilibrium
With the partition of z(`). We can use the following transformation for the land used for
soy (given proposition 1), which is very useful for the structural estimation:
65
∫`∈TSi
z(`)1
1−φd`
T Si=
∫∞zSC
z(`)1
1−φf(z|z > zi)dz(`)
∫`∈TSi
z(`)1
1−φd` = T Si∫∞zSC
z(`)1
1−φf(z|z > zSCi )dz(`)
= P (z > zSCi )Ti∫∞zSC
z(`)1
1−φf(z)
P (z > zSCi )dz(`)
= Ti∫∞zSC
z(`)1
1−φf(z)dz(`)
Using these transformations, we can restate the equilibrium conditions as:
T Si = Ti
∫ ∞zSC
z(`)f(z)dz(`) (44)
TCi = Ti
∫ SC
zCFz(`)f(z)dz(`) (45)
LSi = (pSe
−γSDroadsiφ
wi)
11−φ
∫ ∞zSC
z(`)1
1−φdz(`) (46)
LCi = (pCe
−γCDroadsiφ
wi)
11−φ
∫ SC
zCFz(`)
11−φdz(`) (47)
LUi =αUαM
(LAiφ− T Si
fSwi− TCi
fCwi
)(48)
wi = (V eγαMDroadipαMM
aαUS)
11−αU (49)
pUi =wiaU
(50)
YSi =
(pSe
γSDroadsiφ
wi
) φ1−φ
Ti
∫`∈TSi
z1
1−φd` (51)
YCi =
(pCe
γCDroadsiφ
wi
) φ1−φ
Ti
∫`∈TCi
z1
1−φd` (52)
66
Recommended